Advancing global urban mapping with multimodal robustness and versatile applications

A robust multimodal global urban monitoring approach is vital to understanding the urbanization process and its multidimensional challenges. However, most existing methods presuppose the complete availability of all data sources, leading to performance degradation under missing observations. Furthermore, global urban mapping efforts have focused on using the resulting map, while the potential of the underlying mapping models remains largely unexplored. To address these challenges, this paper proposes a student–teacher framework in which the student model adapts to multi-scenario inputs and learns from contextually ensembled teacher models through a decoupled head, enhancing robustness in multimodal urban mapping. The resulting model, Global Urban Mapper 2.0 (GUM 2.0), is leveraged through its pretrained weights to support various tasks and enable the exploration of model-based applications. The model is trained on a global dataset containing 33,138 multimodal patches (encompassing optical, synthetic-aperture radar (SAR), and topographic data) and validated on a global independent test set. Results show that GUM 2.0 outperforms other global urban mapping products (OA + 2.85%; mIoU + 7.16%) and surpasses existing algorithms across all input scenarios (+0.31%–12.95%). In addition, its pretrained weights enhance performance on downstream tasks, including road extraction, building identification, building height estimation, and Local Climate Zone (LCZ) mapping, and remain effective even with missing modalities. The training algorithm and GUM 2.0 model will be made available at https://github.com/RCAIG-POLYU/GlobalUrbanMapper2.0-, providing robust support for global urban monitoring, diverse urban-related tasks, and contributing to sustainable urban development initiatives.